WO2018049598A1 - 一种眼底图像增强方法及系统 - Google Patents

一种眼底图像增强方法及系统 Download PDF

Info

Publication number
WO2018049598A1
WO2018049598A1 PCT/CN2016/099026 CN2016099026W WO2018049598A1 WO 2018049598 A1 WO2018049598 A1 WO 2018049598A1 CN 2016099026 W CN2016099026 W CN 2016099026W WO 2018049598 A1 WO2018049598 A1 WO 2018049598A1
Authority
WO
WIPO (PCT)
Prior art keywords
sub
image
blood vessel
image block
residual
Prior art date
Application number
PCT/CN2016/099026
Other languages
English (en)
French (fr)
Inventor
吴俊豪
杨烜
裴继红
Original Assignee
深圳大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳大学 filed Critical 深圳大学
Priority to PCT/CN2016/099026 priority Critical patent/WO2018049598A1/zh
Publication of WO2018049598A1 publication Critical patent/WO2018049598A1/zh
Priority to US16/221,781 priority patent/US10748268B2/en

Links

Images

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B3/00—Apparatus for testing the eyes; Instruments for examining the eyes
    • A61B3/0016—Operational features thereof
    • A61B3/0025—Operational features thereof characterised by electronic signal processing, e.g. eye models
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/50—Image enhancement or restoration using two or more images, e.g. averaging or subtraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/70—Denoising; Smoothing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/73—Deblurring; Sharpening
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0012—Biomedical image inspection
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B3/00—Apparatus for testing the eyes; Instruments for examining the eyes
    • A61B3/10—Objective types, i.e. instruments for examining the eyes independent of the patients' perceptions or reactions
    • A61B3/12—Objective types, i.e. instruments for examining the eyes independent of the patients' perceptions or reactions for looking at the eye fundus, e.g. ophthalmoscopes
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20016—Hierarchical, coarse-to-fine, multiscale or multiresolution image processing; Pyramid transform
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20021—Dividing image into blocks, subimages or windows
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20024—Filtering details
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20172—Image enhancement details
    • G06T2207/20192—Edge enhancement; Edge preservation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30041—Eye; Retina; Ophthalmic
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30101—Blood vessel; Artery; Vein; Vascular
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/90—Dynamic range modification of images or parts thereof
    • G06T5/92—Dynamic range modification of images or parts thereof based on global image properties
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/136—Segmentation; Edge detection involving thresholding

Definitions

  • the present invention relates to the field of image processing technologies, and in particular, to a fundus image enhancement method and system.
  • Ocular imaging is an important means of medically assisted diagnosis and treatment. Many eye diseases can be directly or indirectly determined by analyzing eyeball blood vessel images. In the eye image, there are various ocular blood vessels of various thicknesses, and the enhancement of these blood vessels can obtain a clearer and more accurate image of the ocular blood vessels, which is helpful for assisting clinical diagnosis.
  • Field smoothing method That is, the average value of a pixel in the image and the gray level of its neighboring pixels is used as the gray value of the pixel.
  • the advantage of this method is simplicity.
  • the disadvantage is that the image of the blood vessels of the eye is blurred, which greatly reduces the clarity of the blood vessel.
  • Multi-image averaging method In this method, a plurality of eyeball blood vessel images of the same person are averaged.
  • the advantage of this method is that noise can be suppressed to a certain extent.
  • the disadvantage is that multiple eyeball images are needed, which is not suitable for a single fundus image.
  • the image denoising method based on the sparse representation obtains the redundant dictionary through training, and then reconstructs the original image according to the sparse coefficient. Since the selected dictionary atom has no noise, the noise suppression image can be obtained. This method has a better anti-noise effect, but there is still a problem of losing weak weak blood vessels in the application of fundus image enhancement.
  • the technical problem to be solved by the present invention is to provide a fundus image enhancement method and system, which aims to solve the problem that the fundus image enhancement method of the prior art cannot better preserve the weak blood vessels while performing fundus vascular enhancement.
  • the present invention is implemented as follows:
  • a fundus image enhancement method includes the following steps:
  • Step A constructing a blood vessel dictionary by using a fundus learning image, wherein the blood vessel dictionary includes a set number of first sub-image blocks;
  • Step B performing Frangi filtering on the fundus image to be enhanced, and dividing the image obtained by Frangi filtering into a plurality of second sub-image blocks overlapping each other;
  • Step C directional filtering the second sub-image block by using a directional filter, and determining, according to the direction filtering result, whether the fundus blood vessel included in the second sub-image block is a thick blood vessel or a weak blood vessel;
  • Step D determining a blood vessel region in the second sub-image block, and setting a residual weight and a residual of the blood vessel region in the second sub-image block according to a type of a fundus blood vessel included in the second sub-image block Difference threshold
  • Step E omitting the second sub-image block with each first sub-image block in the vascular dictionary, determining a first sub-image block in which the inner product is the largest, and calculating the first inner product maximum a sparse coefficient corresponding to the sub-image block;
  • Step F calculating a residual image by using the first sub-image block and the second sub-image block having the largest inner product, and calculating a blood vessel region in the second sub-image block by using a residual weight of the blood vessel region Residual
  • Step G when the norm of the residual is greater than the residual threshold, set the residual image as the second sub-image block, and jump to step E, otherwise, jump to step H;
  • Step H reconstructing the second sub-image block by using the sparse coefficient
  • Step I reconstructing the fundus image by using each reconstructed second sub-image block, thereby obtaining an enhanced fundus image.
  • step A includes:
  • Step A1 dividing the fundus learning image into a plurality of first sub-image blocks of the same size; the number of the first sub-image blocks is greater than the set number;
  • Step A2 inner product of each of the first sub-image blocks
  • Step A3 selecting the set number of first sub-image blocks with the smallest inner product to construct the blood vessel dictionary.
  • step B includes:
  • Step B1 setting the fundus image to be enhanced to I(x, y), and the two-dimensional Gaussian function of scale ⁇ is G(x, y; ⁇ ), and the fundus image to be enhanced is used by the two-dimensional Gaussian function I(x, y) is smoothed to obtain a smoothed image I ⁇ (x, y):
  • Step B2 at scale ⁇ , calculated smoothed image I ⁇ (x, y) the midpoint (x, y) Hessian matrix H ⁇ (x, y) at:
  • Step B3 Perform eigenvalue analysis on the Hessian matrix H ⁇ (x, y) to obtain eigenvalues ⁇ 1 , ⁇ 2 ,
  • ; the vascular characteristics at the scale s are:
  • ⁇ and C are preset constants
  • Step B4 At multiple scales, take the maximum value of ⁇ 0 (s) at each scale as the Frangi filtering result v of the fundus image I(x, y) to be enhanced:
  • s min and s max are the minimum scale and the largest scale, respectively;
  • Step B5 The Frangi filtering result v is divided into a plurality of second sub-image blocks that overlap each other.
  • step C includes:
  • Step C2 Assuming that the vascular region in the direction filter of direction ⁇ i is ⁇ 1 and the non-vascular region is ⁇ 2 , the energy of each of the two regions is calculated.
  • N 1 is the number of pixels in ⁇ 1
  • N 2 is the number of pixels in ⁇ 2 ;
  • Step C3 Calculation versus Energy difference:
  • Step C4 Determine the maximum energy difference among the above eight directions:
  • Step C5 Determine the blood vessel type according to the E max . If E max ⁇ T, the fundus image included in the second sub-image block is a thick blood vessel, otherwise it is a weak blood vessel.
  • step D includes:
  • Step D1 taking the blood vessel region in the directional filter corresponding to the largest energy difference among the eight directions as the actual region ⁇ 1 of the blood vessel, and directional filter corresponding to the largest energy difference among the eight directions The non-vascular region as the non-vascular actual region ⁇ 2 ;
  • step E includes:
  • Step E1 vectorizing the second sub-image block into x, and the i-th first sub-image block in the vascular dictionary is d i ;
  • Step E2 taking the largest inner product block of the first sub-image block and the second sub-image block x in the blood vessel dictionary as the selected first first sub-image block d r0 :
  • k is the number of the first sub-image block in the vascular dictionary
  • r 0 is the index number of the dictionary
  • ⁇ x, d i > is an inner product operation of x and d i ;
  • Step E3 calculating a first sub-image blocks corresponding to d r0 sparse coefficients ⁇ r0:
  • ⁇ r0 ⁇ x, d r0 >.
  • step F includes:
  • Step F1 calculating a residual image R of the blood vessel region in the second sub-image block:
  • R x- ⁇ x, d r0 >d r0 ;
  • Step F2 Multiplying the residual R by the residual weight of the blood vessel region in the second sub-image block to obtain the final residual of the blood vessel region in the second sub-image block.
  • the reconstructed second sub-image block is:
  • S is a plurality of sets of sparse coefficients determined in step E
  • d r0 is the first sub-image block with the largest inner product determined by performing step E
  • ⁇ r0 is the sparse coefficient corresponding to d r0 .
  • step I includes:
  • a fundus image enhancement system comprising:
  • vascular dictionary construction module that constructs a vascular dictionary using a fundus learning image, the vascular dictionary including a set number of first sub-image blocks;
  • An image filtering and dividing module which performs Frangi filtering on the fundus image to be enhanced, and divides the image obtained by Frangi filtering into a plurality of second sub-image blocks overlapping each other;
  • a blood vessel type judging module that performs direction filtering on the second sub-image block by using a directional filter, and determines, according to the direction filtering result, whether the fundus blood vessel included in the second sub-image block is a thick blood vessel or a weak blood vessel;
  • a blood vessel region and its residual weight and residual threshold determination module that determines a blood vessel region in the second sub-image block and sets the second sub-portion according to a type of fundus blood vessel included in the second sub-image block The residual weight and residual threshold of the blood vessel region in the image block;
  • a sparse coefficient calculation module which integrates the second sub-image block with each first sub-image block in the vascular dictionary, determines a first sub-image block in which the inner product is the largest, and calculates the inner product maximum a sparse coefficient corresponding to the first sub-image block;
  • a blood vessel region residual calculation module that calculates a residual image using the first sub-image block and the second sub-image block having the largest inner product, and calculates the second sub image using residual weights of the blood vessel region The residual of the vascular region in the block;
  • a jump module when the norm of the residual is greater than the residual threshold, setting a residual image as a second sub-image block, and jumping to the sparse coefficient calculation module; otherwise, jumping to a second sub-image block reconstruction module;
  • a second sub-image block reconstruction module that reconstructs the second sub-image block by using the sparse coefficient
  • a fundus image reconstruction module reconstructs the fundus image using each reconstructed second sub-image block to obtain an enhanced fundus image.
  • vascular dictionary building module includes:
  • a fundus learning image dividing module which divides the fundus learning image into a plurality of first sub-image blocks of the same size; the number of the first sub-image blocks is greater than the set number;
  • the vascular dictionary construction sub-module selects the set number of first sub-image blocks with the smallest inner product to construct the vascular dictionary.
  • the image filtering and dividing module includes:
  • a smoothing filter module wherein the fundus image to be enhanced is I(x, y), and the two-dimensional Gaussian function of scale ⁇ is G(x, y; ⁇ ), and the two-dimensional Gaussian function is used to The fundus image I(x, y) is smoothed to obtain a smoothed image I ⁇ (x, y):
  • the Hessian matrix calculation module calculates the Hessian matrix H ⁇ (x, y) at the point (x, y) of the smoothed image I ⁇ (x, y) at the scale ⁇ :
  • An eigenvalue analysis module that performs eigenvalue analysis on the Hessian matrix H ⁇ (x, y) to obtain eigenvalues ⁇ 1 , ⁇ 2 ,
  • ; the vascular characteristics at the scale s are:
  • ⁇ and C are preset constants
  • the Frangi filtering result generating module takes the maximum value of ⁇ 0 (s) at each scale as the Frangi filtering result v of the fundus image I(x, y) to be enhanced:
  • s min and s max are the minimum scale and the largest scale, respectively;
  • a second sub-image dividing module that divides the Frangi filtering result v into a plurality of second sub-image blocks that overlap each other.
  • the blood vessel type determining module includes:
  • An energy calculation module that assumes that the vascular region in the directional direction of the ⁇ i is ⁇ 1 and the non-vascular region is ⁇ 2 , and calculates the energy of each of the two regions.
  • N 1 is the number of pixels in ⁇ 1
  • N 2 is the number of pixels in ⁇ 2 ;
  • a maximum energy difference determination module that determines the largest energy difference among the above eight directions:
  • a blood vessel type judging sub-module which determines a blood vessel type according to the E max , and if E max ⁇ T, the fundus image included in the second sub-image block is a thick blood vessel, and otherwise is a weak blood vessel.
  • the blood vessel region and its residual weight and residual threshold determination module include:
  • a blood vessel region determining module that uses a blood vessel region in a directional filter corresponding to a maximum energy difference among the eight directions as a blood vessel actual region ⁇ 1 , and a direction corresponding to a maximum energy difference among the eight directions
  • the area of the blood vessel in the filter acts as the actual area of the blood vessel ⁇ 2 ;
  • the sparse coefficient calculation module includes:
  • An image vector module that vectorizes the second sub-image block into x, and the i-th first sub-image block in the vascular dictionary is d i ;
  • a first sub-image block selection module which uses the innermost product of each first sub-image block and the second sub-image block x in the vascular dictionary as the selected first first sub-image block d r0 :
  • k is the number of the first sub-image block in the vascular dictionary
  • r 0 is the index number of the dictionary
  • ⁇ x, d i > is an inner product operation of x and d i ;
  • Sparse coefficient calculation sub-module that calculates a first image sub-block corresponding to d r0 sparse coefficients ⁇ r0:
  • ⁇ r0 ⁇ x, d r0 >.
  • the blood vessel region residual calculation module includes:
  • a residual initial calculation module that calculates a residual image R of a blood vessel region in the second sub-image block:
  • R x- ⁇ x, d r0 >d r0 ;
  • a residual weighting module that multiplies the residual R by the residual weight of the blood vessel region in the second sub-image block to obtain a final residual of the blood vessel region in the second sub-image block.
  • the reconstructed second sub-image block is:
  • S is a set of a plurality of sparse coefficients determined by the sparse coefficient calculation module a plurality of times
  • d r0 is a first sub-image block having the largest inner product determined by the sparse coefficient calculation module
  • ⁇ r0 is d The sparse coefficient corresponding to r0 .
  • the fundus image reconstruction module is specifically configured to:
  • the invention constructs a blood vessel dictionary by using a fundus learning image; classifies the blood vessels in the second sub-image block into a thick blood vessel and a weak blood vessel by using direction filtering, and sets a residual weight and a residual threshold value of the blood vessel region for the crude blood vessel and the weak blood vessel;
  • the second sub-image block is internally coded with each first sub-image block in the dictionary, the first sub-image block having the largest inner product is selected, and the corresponding sparse coefficient is calculated; and the selected first sub-image block and the blood vessel region are used.
  • the residual weight calculates a residual of the blood vessel region in the second sub-image block, and if the residual is greater than the residual threshold, repeats the process of selecting the first sub-image block and calculating the residual; reconstructing the second sub-image block by using the sparse coefficient, And reconstructing each reconstructed second sub-image block to obtain an enhanced fundus image.
  • the invention reduces the phenomenon of background noise and weak blood vessel loss caused by blood vessel enhancement by Frangi filtering in the prior art, realizes enhancement of fundus image, improves visual effect of fundus image, and can be used for pretreatment of fundus image analysis.
  • FIG. 1 is a schematic overall flow chart of a fundus image enhancement method provided by the present invention
  • FIG. 2 is a schematic diagram showing the overall composition of a fundus image enhancement system provided by the present invention.
  • Figure 3 Schematic diagram of the direction of the eight directional filters.
  • the fundus image enhancement method provided by the present invention comprises the following steps:
  • Step A Constructing a blood vessel dictionary using a fundus learning image, the blood vessel dictionary including a set number of first sub-image blocks.
  • Step A specifically includes:
  • Step A1 dividing the fundus learning image into a plurality of first sub-image blocks of the same size, the number of the first sub-image blocks Should be greater than the set number.
  • the fundus learning image is segmented into a plurality of first sub-image blocks of 8*8 size, and the first sub-image block should include fundus image features such as thick blood vessels, thin blood vessels, and highlights. .
  • Step A2 inner product of each of the first sub-image blocks. The smaller the inner product, the smaller the similarity between the two first sub-image blocks.
  • Step A3 Select a set number of first sub-image blocks with the smallest inner product to construct a blood vessel dictionary. Assuming that the set number is K, the most dissimilar K sub-image blocks are selected to form a vascular dictionary.
  • Step B Frangi filtering is performed on the fundus image to be enhanced, and the image obtained by Frangi filtering is divided into a plurality of second sub-image blocks overlapping each other.
  • Step B specifically includes:
  • Step B1 Let the fundus image to be enhanced be I(x, y), the two-dimensional Gaussian function of scale ⁇ is G(x, y; ⁇ ), and the fundus image I(x, y) to be enhanced by the two-dimensional Gaussian function Smoothing is performed to obtain a smoothed image I ⁇ (x, y):
  • Step B2 at scale ⁇ , calculated smoothed image I ⁇ (x, y) the midpoint (x, y) Hessian matrix H ⁇ (x, y) at:
  • Step B3 Perform eigenvalue analysis on the Hessian matrix H ⁇ (x, y) to obtain eigenvalues ⁇ 1 , ⁇ 2 ,
  • ⁇ and C are preset constants.
  • Step B4 At multiple scales, the maximum value of ⁇ 0 (s) at each scale is taken as the Frangi filtering result of the fundus image I(x, y) to be enhanced v:
  • s min and s max are the minimum scale and the largest scale, respectively.
  • Step B5 The Frangi filtering result v is divided into a plurality of second sub-image blocks that overlap each other.
  • Step C directional filtering is performed on the second sub-image block by using a directional filter, and determining whether the fundus blood vessel included in the second sub-image block is a thick blood vessel or a weak blood vessel according to the direction filtering result.
  • Step C specifically includes:
  • Step C2 Assuming that the vascular region (white region) in the direction filter of direction ⁇ i is ⁇ 1 and the non-vascular region (black region) is ⁇ 2 , the energy of each of the two regions is calculated.
  • N 1 is the number of pixels in ⁇ 1
  • N 2 is the number of pixels in ⁇ 2 .
  • Step C3 Calculation versus Energy difference:
  • Step C4 Determine the maximum energy difference among the above eight directions:
  • Step C5 Determine the blood vessel type according to E max . If E max ⁇ T, the fundus image contained in the second sub-image block is a thick blood vessel, otherwise it is a weak blood vessel. T is the default value.
  • Step D determining a blood vessel region in the second sub-image block, and setting a residual weight and a residual threshold of the blood vessel region in the second sub-image block according to the type of the fundus blood vessel included in the second sub-image block.
  • Step D specifically includes:
  • Step D1 The vascular region in the directional filter corresponding to the largest energy difference among the eight directions (even if the directional filter corresponding to ⁇ i of E max is the largest) is taken as the actual region ⁇ 1 of the blood vessel, and will be in eight directions.
  • the non-vascular region in the directional filter corresponding to the largest energy difference (even if the directional filter corresponding to ⁇ i of E max is the largest) is the non-vascular actual region ⁇ 2 .
  • Step D may also include:
  • Step E omitting the second sub-image block with each first sub-image block in the vascular dictionary, determining the first sub-image block in which the inner product is the largest, and calculating the sparsity corresponding to the first sub-image block having the largest inner product. coefficient.
  • Step E specifically includes:
  • Step E1 Vectorizing the second sub-image block into x, and the i-th first sub-image block in the vascular dictionary is d i .
  • Step E2 taking the largest inner product block of each of the first sub-image block and the second sub-image block x in the vascular dictionary as the selected first first sub-image block d r0 :
  • k is the number of the first sub-image block in the vascular dictionary
  • r 0 is the index number of the dictionary
  • ⁇ x, d i > is the inner product operation of x and d i .
  • Step E3 calculating a first sub-image blocks corresponding to d r0 sparse coefficients ⁇ r0:
  • Step F calculating a residual image by using the first sub-image block and the second sub-image block having the largest inner product, and calculating the residual of the blood vessel region in the second sub-image block by using the residual weight of the blood vessel region.
  • Step F specifically includes:
  • Step F1 Calculating the residual image R of the blood vessel region in the second sub-image block:
  • R x- ⁇ x, d r0 >d r0 ;
  • Step F2 Multiplying the residual R by the residual weight of the blood vessel region in the second sub-image block to obtain the final residual of the blood vessel region in the second sub-image block.
  • Step G When the norm of the residual is greater than the residual threshold, set the residual image as the second sub-image block and jump to step E, otherwise, jump to step H. That is, if the norm
  • Step H reconstructing the second sub-image block with the sparse coefficient.
  • the reconstructed second sub-image block is:
  • S is a plurality of sets of sparse coefficients determined in step E
  • d r0 is the first sub-image block with the largest inner product determined by performing step E
  • ⁇ r0 is the sparse coefficient corresponding to d r0 .
  • Step I Reconstructing the fundus image with each reconstructed second sub-image block, thereby obtaining an enhanced fundus image.
  • Step I includes:
  • the present invention further provides a fundus image enhancement system, comprising: a vascular dictionary construction module 1, an image filtering and division module 2, a blood vessel type determination module 5, a blood vessel region and The residual weight and residual threshold determination module 4, the sparse coefficient calculation module 3, the blood vessel region residual calculation module 6, the jump module 7, the second sub-image block reconstruction module 8, and the fundus image reconstruction module 9.
  • a fundus image enhancement system comprising: a vascular dictionary construction module 1, an image filtering and division module 2, a blood vessel type determination module 5, a blood vessel region and The residual weight and residual threshold determination module 4, the sparse coefficient calculation module 3, the blood vessel region residual calculation module 6, the jump module 7, the second sub-image block reconstruction module 8, and the fundus image reconstruction module 9.
  • the vascular dictionary construction module 1 constructs a vascular dictionary using a fundus learning image including a set number of first sub-image blocks.
  • the vascular dictionary construction module 1 includes a fundus learning image dividing module, a first sub-image block inner product module, and a blood vessel dictionary construction sub-module.
  • the fundus learning image dividing module divides the fundus learning image into a plurality of first sub-image blocks of the same size, the first sub-picture The number of image blocks is greater than the set number.
  • the vascular dictionary construction sub-module selects a set number of first sub-image blocks with the smallest inner product to construct a vascular dictionary.
  • the image filtering and dividing module 2 performs Frangi filtering on the fundus image to be enhanced, and divides the image obtained by Frangi filtering into a plurality of second sub-image blocks overlapping each other.
  • the image filtering and dividing module 2 includes a smoothing filtering module, a Hessian matrix computing module, an eigenvalue analysis module, a Frangi filtering result generating module, and a second sub-image dividing module.
  • the smoothing filter module sets the fundus image to be enhanced to I(x, y), the two-dimensional Gaussian function of scale ⁇ to G(x, y; ⁇ ), and the fundus image I(x, y) to be enhanced by the two-dimensional Gaussian function. Smoothing is performed to obtain a smoothed image I ⁇ (x, y):
  • the Hessian matrix calculation module calculates the Hessian matrix H ⁇ (x, y) at the point (x, y) of the smoothed image I ⁇ (x, y) at the scale ⁇ :
  • the eigenvalue analysis module performs eigenvalue analysis on the Hessian matrix H ⁇ (x, y) to obtain eigenvalues ⁇ 1 , ⁇ 2 ,
  • ; the vascular characteristics at the scale s are:
  • ⁇ and C are preset constants.
  • the Frangi filtering result generation module takes the maximum value of ⁇ 0 (s) at each scale as the Frangi filtering result v of the fundus image I(x, y) to be enhanced:
  • s min and s max are the minimum scale and the largest scale, respectively.
  • the second sub-image dividing module divides the Frangi filtering result v into a plurality of mutually overlapping second sub-image blocks.
  • the blood vessel type judging module 5 performs direction filtering on the second sub-image block by using the directional filter, and determines whether the fundus blood vessel included in the second sub-image block is a thick blood vessel or a weak blood vessel according to the direction filtering result.
  • the blood vessel type judging module 5 includes a directional filter setting module, an energy calculation module, an energy difference calculation module, a maximum energy difference determination module, and a blood vessel type judgment sub-module.
  • the energy calculation module assumes that the vascular region in the direction filter of direction ⁇ i is ⁇ 1 and the non-vascular region is ⁇ 2 , and the energy of each region is calculated.
  • N 1 is the number of pixels in ⁇ 1
  • N 2 is the number of pixels in ⁇ 2 .
  • the maximum energy difference determination module determines the largest energy difference among the above eight directions:
  • the blood vessel type judging sub-module determines the blood vessel type according to E max . If E max ⁇ T, the fundus image contained in the second sub-image block is a thick blood vessel, otherwise it is a weak blood vessel.
  • the blood vessel region and its residual weight and residual threshold determination module 4 determines a blood vessel region in the second sub-image block, and sets a blood vessel region in the second sub-image block according to a type of fundus blood vessel included in the second sub-image block. Residual weight and residual threshold.
  • the vessel region and its residual weight and residual threshold determination module 4 includes a vessel region determination module and a residual weight and residual threshold determination module.
  • the blood vessel region determining module uses the blood vessel region in the directional filter corresponding to the largest energy difference among the eight directions as the actual region ⁇ 1 of the blood vessel, and the blood vessel region in the directional filter corresponding to the largest energy difference among the eight directions As the actual area of the blood vessel ⁇ 2 .
  • the sparse coefficient calculation module 3 integrates the second sub-image block with each first sub-image block in the vascular dictionary, determines the first sub-image block in which the inner product is the largest, and calculates the first sub-image block corresponding to the largest inner product. Sparse coefficient.
  • the sparse coefficient calculation module 3 includes an image vector module, a first sub-image block selection module, and a sparse coefficient calculation sub-module.
  • the image vector module vectorizes the second sub-image block into x, and the i-th first sub-image block in the vascular dictionary is d i .
  • the first sub-image block selection module takes the largest inner product block of each of the first sub-image block and the second sub-image block x in the blood vessel dictionary as the selected first first sub-image block d r0 :
  • k is the number of the first sub-image block in the vascular dictionary
  • r 0 is the index number of the dictionary
  • ⁇ x, d i > is the inner product operation of x and d i .
  • Sparse coefficient calculation module calculates a first sub-block sub-image corresponding to d r0 sparse coefficients ⁇ r0:
  • the blood vessel region residual calculation module 6 calculates a residual image using the first sub-image block and the second sub-image block having the largest inner product, and calculates the residual of the blood vessel region in the second sub-image block by using the residual weight of the blood vessel region.
  • the vascular region residual calculation module 6 includes a residual initial calculation module and a residual weighting module.
  • the residual initial calculation module calculates a residual image R of the blood vessel region in the second sub-image block:
  • R x- ⁇ x, d r0 >d r0 ;
  • the residual weighting module multiplies the residual R by the residual weight weighted summation of the blood vessel regions in the second sub-image block as the final residual of the blood vessel region in the second sub-image block.
  • the jump module 7 sets the residual image as the second sub-image block when the norm of the residual is greater than the residual threshold, and jumps to the sparse coefficient calculation module 3, otherwise, jumps to the second sub-image block.
  • the second sub-image block reconstruction module 8 reconstructs the second sub-image block using the sparse coefficients.
  • the reconstructed second sub-image block is:
  • S is a set of a plurality of sparse coefficients determined by the sparse coefficient calculation module a plurality of times
  • d r0 is a first sub-image block having the largest inner product determined by the sparse coefficient calculation module
  • ⁇ r0 is d The sparse coefficient corresponding to r0 .
  • the fundus image reconstruction module 9 reconstructs the fundus image using each of the reconstructed second sub-image blocks, thereby obtaining an enhanced fundus image.
  • the fundus image reconstruction module 9 is specifically used to:
  • each module in the system can refer to the corresponding steps in the aforementioned fundus image enhancement method.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Medical Informatics (AREA)
  • General Health & Medical Sciences (AREA)
  • Animal Behavior & Ethology (AREA)
  • Signal Processing (AREA)
  • Molecular Biology (AREA)
  • Surgery (AREA)
  • Biomedical Technology (AREA)
  • Ophthalmology & Optometry (AREA)
  • Public Health (AREA)
  • Veterinary Medicine (AREA)
  • Biophysics (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Quality & Reliability (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Image Processing (AREA)
  • Eye Examination Apparatus (AREA)
  • Image Analysis (AREA)

Abstract

一种眼底图像增强方法及系统，利用眼底学习图像构建血管字典；对待增强眼底图像进行Frangi滤波，利用方向滤波将第二子图像块中的血管分为粗血管和细弱血管，并据此设定血管区域的残差权重及残差阈值；将第二子图像块与字典中各第一子图像块内积，选取内积最大的第一子图像块，并计算其相应稀疏系数；利用选取的第一子图像块和第二子图像块计算残差图像，并利用血管区域的残差权重计算第二子图像块中血管区域的残差，如果残差大于残差阈值，则将残差图像设置为第二子图像块，并重复选取第一子图像块与计算残差过程；利用稀疏系数重构第二子图像块，并将各重构的第二子图像块重组，得到增强的眼底图像，可有效抑制背景噪声，保留细弱血管。

Description

一种眼底图像增强方法及系统 技术领域
本发明涉及图像处理技术领域，尤其涉及一种眼底图像增强方法及系统。
背景技术
眼部成像是医学辅助诊疗的重要手段，通过分析眼球血管图像可以直接或间接地判断出许多眼部疾病。在眼部图像中，存在各种不同粗细程度的眼部血管，增强这些血管，可以得到更清晰准确的眼部血管图像，有利于辅助临床诊断。
眼底图像增强方法有很多，一般常用的方法有：
领域平滑法。即利用图像中某一像素以及它的邻域像素灰度的平均值作为该像素的灰度值。该方法的优点是简单，缺点是会使得眼部血管图像变得模糊，大大降低了血管的清晰度。
保存边界平滑法。即设计不同模板，计算图像中某一像素点所处邻域像素灰度的方差，将方差最小的模板所含像素的灰度平均值作为该像素点的灰度值。这种方法的优点是可以较好地保存边界，缺点是眼底图像中目标是线结构，难以通过方差区分噪声与目标。
多图像平均法。此种方法是取同一人的多幅眼球血管图像进行平均处理。该方法的优点是可一定成程度抑制噪声，缺点是需要多张眼球血管图像，不适用于单张眼底图像。
Frangi滤波图像增强。这种方法利用线结构的Hessian矩阵的特征向量方向和特征值，对线结构进行增强，但这类方法会使细弱的弱血管丢失。
基于稀疏表示的图像去噪方法通过训练得到冗余字典，再根据稀疏系数重构原图像，由于选取的字典原子没有噪声，从而可以得到抑噪图像。这种方法具有较好的抑噪效果，但是在应用到眼底图像增强问题中仍存在使细弱的弱血管丢失的问题。
由此可见，现有的眼底图像增强方法都无法在进行眼底血管增强的同时较好地保留细弱血管。
发明内容
本发明所要解决的技术问题是，提供一种眼底图像增强方法及系统，旨在解决现有技术的眼底图像增强方法无法在进行眼底血管增强的同时较好地保留细弱血管的问题。本发明是这样实现的：
一种眼底图像增强方法，包括如下步骤：
步骤A：利用眼底学习图像构建血管字典，所述血管字典中包括设定数量的第一子图像块；
步骤B：对待增强的眼底图像进行Frangi滤波，并将Frangi滤波得到的图像划分为若干相互重叠的第二子图像块；
步骤C：利用方向滤波器对所述第二子图像块进行方向滤波，并根据方向滤波结果判断所述第二子图像块中包含的眼底血管是粗血管还是细弱血管；
步骤D：确定所述第二子图像块中的血管区域，并根据所述第二子图像块中包含的眼底血管的类型设置所述第二子图像块中的血管区域的残差权重和残差阈值；
步骤E：将所述第二子图像块与所述血管字典中的各第一子图像块内积，确定出其中内积最大的第一子图像块，并计算所述内积最大的第一子图像块对应的稀疏系数；
步骤F：利用所述内积最大的第一子图像块和所述第二子图像块计算残差图像，并利用所述血管区域的残差权重计算所述第二子图像块中血管区域的残差；
步骤G：当所述残差的范数大于所述残差阈值时，将残差图像设置为第二子图像块，并跳转至步骤E，否则，跳转至步骤H；
步骤H：利用所述稀疏系数重构所述第二子图像块；
步骤I：利用各重构的第二子图像块重构所述眼底图像，从而得到增强的眼底图像。
进一步地，所述步骤A包括：
步骤A1：将所述眼底学习图像分割成若干大小相同的第一子图像块；所述第一子图像块的数量大于所述设定数量；
步骤A2：将各第一子图像块两两进行内积；
步骤A3：选取内积最小的所述设定数量个第一子图像块构建所述血管字典。
进一步地，所述步骤B包括：
步骤B1：设待增强的眼底图像为I(x，y)，尺度为σ的二维高斯函数为G(x，y；σ)，利用所述二维高斯函数对所述待增强的眼底图像I(x，y)进行平滑处理，得到平滑图像Iσ(x，y)：
Figure PCTCN2016099026-appb-000001
其中，
Figure PCTCN2016099026-appb-000002
为卷积操作；
步骤B2：在尺度σ下，计算平滑图像Iσ(x，y)中点(x，y)处的Hessian矩阵Hσ(x，y)：
Figure PCTCN2016099026-appb-000003
步骤B3：对所述Hessian矩阵Hσ(x，y)做特征值分析，得到特征值λ1、λ2，|λ1|＜|λ2|；尺度s下的血管特征为：
Figure PCTCN2016099026-appb-000004
其中，
Figure PCTCN2016099026-appb-000005
β和C是预设常数；
步骤B4：在多尺度下，取各尺度下υ0(s)的最大值作为所述待增强的眼底图像I(x，y)的Frangi滤波结果v：
Figure PCTCN2016099026-appb-000006
其中，smin和smax分别是最小尺度和最大尺度；
步骤B5：将所述Frangi滤波结果v划分为若干相互重叠的第二子图像块。
进一步地，所述步骤C包括：
步骤C1：设置方向分别为θ1＝0，
Figure PCTCN2016099026-appb-000007
Figure PCTCN2016099026-appb-000008
的8个方向滤波器；
步骤C2：假设方向为θi的方向滤波器中血管区域为Ω1，非血管区域为Ω2，计算两个区域各自的能量
Figure PCTCN2016099026-appb-000009
和
Figure PCTCN2016099026-appb-000010
Figure PCTCN2016099026-appb-000011
Figure PCTCN2016099026-appb-000012
其中v(x，y)是Frangi滤波结果v在(x，y)的值，N1是Ω1中像素个数，N2是Ω2中像素个数；
步骤C3：计算
Figure PCTCN2016099026-appb-000013
与
Figure PCTCN2016099026-appb-000014
的能量差：
Figure PCTCN2016099026-appb-000015
步骤C4：确定上述8个方向中最大的能量差：
Figure PCTCN2016099026-appb-000016
步骤C5：根据所述Emax判断血管类型，如果Emax≥T，则所述第二子图像块中包含的眼底图像为粗血管，否则为细弱血管。
进一步地，所述步骤D包括：
步骤D1：将所述8个方向中最大的能量差所对应的方向滤波器中的血管区域作为血管实际的区域Ω1，将所述8个方向中最大的能量差所对应的方向滤波器中的非血管区域作为非血管实际的区域Ω2；
步骤D2：对于包含的眼底图像为粗血管的第二子图像块，将其血管区域Ω1的残差权重设为1，残差阈值TR＝T1；对于包含的眼底图像为细弱血管的第二子图像块，将其血管区域Ω1的残差权重设为1/υmax，残差阈值TR＝T2，其中υmax是该第二子图像块Frangi滤波结果的最大值。
进一步地，所述步骤E包括：
步骤E1：将所述第二子图像块向量化为x，所述血管字典中第i个第一子图像块为di；
步骤E2：将所述血管字典中各第一子图像块与所述第二子图像块x内积最大者作为选中的第一个第一子图像块dr0：
Figure PCTCN2016099026-appb-000017
其中，k为所述血管字典中第一子图像块的个数，r0是字典的索引号，＜x，di＞是x与di的内积运算；
步骤E3：计算第一子图像块dr0对应的稀疏系数αr0：
αr0＝＜x，dr0＞。
进一步地，所述步骤F包括：
步骤F1：计算所述第二子图像块中血管区域的残差图像R：
R＝x-＜x，dr0＞dr0；
步骤F2：将所述残差R乘以该第二子图像块中血管区域的残差权重加权求和，作为第二子图像块中血管区域的最终残差。
进一步地，重构的所述第二子图像块为：
Figure PCTCN2016099026-appb-000018
其中，S是多次执行步骤E确定出的多个稀疏系数的集合，dr0是每一次执行步骤E确定出的内积最大的第一子图像块，αr0是dr0对应的稀疏系数。
进一步地，所述步骤I包括：
将所有重构的第二子图像块的不相交的部分合并，得到完整的增强的眼底图像。
一种眼底图像增强系统，包括：
血管字典构建模块，其利用眼底学习图像构建血管字典，所述血管字典中包括设定数量的第一子图像块；
图像滤波及划分模块，其对待增强的眼底图像进行Frangi滤波，并将Frangi滤波得到的图像划分为若干相互重叠的第二子图像块；
血管类型判断模块，其利用方向滤波器对所述第二子图像块进行方向滤波，并根据方向滤波结果判断所述第二子图像块中包含的眼底血管是粗血管还是细弱血管；
血管区域及其残差权重和残差阈值确定模块，其确定所述第二子图像块中的血管区域，并根据所述第二子图像块中包含的眼底血管的类型设置所述第二子图像块中的血管区域的残差权重和残差阈值；
稀疏系数计算模块，其将所述第二子图像块与所述血管字典中的各第一子图像块内积，确定出其中内积最大的第一子图像块，并计算所述内积最大的第一子图像块对应的稀疏系数；
血管区域残差计算模块，其利用所述内积最大的第一子图像块和所述第二子图像块计算残差图像，并利用所述血管区域的残差权重计算所述第二子图像块中血管区域的残差；
跳转模块，其在当所述残差的范数大于所述残差阈值时，将残差图像设置为第二子图像块，并跳转至所述稀疏系数计算模块，否则，跳转至第二子图像块重构模块；
第二子图像块重构模块，其利用所述稀疏系数重构所述第二子图像块；
眼底图像重构模块，其利用各重构的第二子图像块重构所述眼底图像，从而得到增强的眼底图像。
进一步地，所述血管字典构建模块包括：
眼底学习图像划分模块，其将所述眼底学习图像分割成若干大小相同的第一子图像块；所述第一子图像块的数量大于所述设定数量；
第一子图像块内积模块，其将各第一子图像块两两进行内积；
血管字典构建子模块，其选取内积最小的所述设定数量个第一子图像块构建所述血管字典。
进一步地，所述图像滤波及划分模块包括：
平滑滤波模块，其设待增强的眼底图像为I(x，y)，尺度为σ的二维高斯函数为G(x，y；σ)，利用所述二维高斯函数对所述待增强的眼底图像I(x，y)进行平滑处理得到平滑图像Iσ(x，y)：
Figure PCTCN2016099026-appb-000019
其中，
Figure PCTCN2016099026-appb-000020
为卷积操作；
Hessian矩阵计算模块，其在尺度σ下，计算平滑图像Iσ(x，y)中点(x，y)处的Hessian矩阵Hσ(x，y)：
Figure PCTCN2016099026-appb-000021
特征值分析模块，其对所述Hessian矩阵Hσ(x，y)做特征值分析，得到特征值λ1、λ2，|λ1|＜|λ2|；尺度s下的血管特征为：
Figure PCTCN2016099026-appb-000022
其中，
Figure PCTCN2016099026-appb-000023
β和C是预设常数；
Frangi滤波结果生成模块，其取各尺度下υ0(s)的最大值作为所述待增强的眼底图像I(x，y)的Frangi滤波结果v：
Figure PCTCN2016099026-appb-000024
其中，smin和smax分别是最小尺度和最大尺度；
第二子图像划分模块，其将所述Frangi滤波结果v划分为若干相互重叠的第二子图像块。
进一步地，所述血管类型判断模块包括：
方向滤波器设置模块，其设置方向分别为θ1＝0，
Figure PCTCN2016099026-appb-000025
Figure PCTCN2016099026-appb-000026
的8个方向滤波器；
能量计算模块，其假设方向为θi的方向滤波器中血管区域为Ω1，非血管区域为Ω2，计算两个区域各自的能量
Figure PCTCN2016099026-appb-000027
和
Figure PCTCN2016099026-appb-000028
Figure PCTCN2016099026-appb-000029
Figure PCTCN2016099026-appb-000030
其中v(x，y)是Frangi滤波结果v在(x，y)的值，N1是Ω1中像素个数，N2是Ω2中像素个数；
能量差计算模块，其计算
Figure PCTCN2016099026-appb-000031
与
Figure PCTCN2016099026-appb-000032
的能量差：
Figure PCTCN2016099026-appb-000033
最大能量差确定模块，其确定上述8个方向中最大的能量差：
Figure PCTCN2016099026-appb-000034
血管类型判断子模块，其根据所述Emax判断血管类型，如果Emax≥T，则所述第二子图像块中包含的眼底图像为粗血管，否则为细弱血管。
进一步地，所述血管区域及其残差权重和残差阈值确定模块包括：
血管区域确定模块，其将所述8个方向中最大的能量差所对应的方向滤波器中的血管区域作为血管实际的区域Ω1，将所述8个方向中最大的能量差所对应的方向滤波器中的血管区域作为血管实际的区域Ω2；
残差权重和残差阈值确定模块，其将包含的眼底图像为粗血管的第二子图像块的血管区域Ω1的残差权重设为1，残差阈值TR＝T1；将包含的眼底图像为细弱血管的第二子图像块的血管区域Ω1的残差权重设为1/υmax，残差阈值TR＝T2，其中υmax是该第二子图像块Frangi滤波结果的最大值。
进一步地，所述稀疏系数计算模块包括：
图像向量模块，其将所述第二子图像块向量化为x，所述血管字典中第i个第一子图像块为di；
第一子图像块选择模块，其将所述血管字典中各第一子图像块与所述第二子图像块x内积最大者作为选中的第一个第一子图像块dr0：
Figure PCTCN2016099026-appb-000035
其中，k为所述血管字典中第一子图像块的个数，r0是字典的索引号，＜x，di＞是x与di的内积运算；
稀疏系数计算子模块，其计算第一子图像块dr0对应的稀疏系数αr0：
αr0＝＜x，dr0＞。
进一步地，所述血管区域残差计算模块包括：
残差初算模块，其计算所述第二子图像块中血管区域的残差图像R：
R＝x-＜x，dr0＞dr0；
残差加权模块，其将所述残差R乘以该第二子图像块中血管区域的残差权重加权求和，作为第二子图像块中血管区域的最终残差。
进一步地，重构的所述第二子图像块为：
Figure PCTCN2016099026-appb-000036
其中，S是所述稀疏系数计算模块多次确定出的多个稀疏系数的集合，dr0是所述稀疏系数计算模块每一次确定出的内积最大的第一子图像块，αr0是dr0对应的稀疏系数。
进一步地，所述眼底图像重构模块具体用于：
将所有重构的第二子图像块的不相交的部分合并，得到完整的增强的眼底图像。
本发明利用眼底学习图像构建血管字典；利用方向滤波将第二子图像块中的血管分类为粗血管和细弱血管，针对粗血管和细弱血管设定血管区域的残差权重及残差阈值；将第二子图像块与字典中的各第一子图像块进行内积，选取内积最大的第一子图像块，并计算其相应的稀疏系数；利用选取的第一子图像块和血管区域的残差权重计算第二子图像块中血管区域的残差，如果残差大于残差阈值，则重复选取第一子图像块与计算残差的过程；利用稀疏系数重构第二子图像块，并将各重构第二子图像块重组，得到增强的眼底图像。本发明减少了现有技术中通过Frangi滤波进行血管增强带来背景噪声和细弱血管丢失的现象，实现了眼底图像的增强，改善了眼底图像视觉效果，可用于眼底图像分析的预处理。
附图说明
图1：本发明提供的眼底图像增强方法的总体流程示意图；
图2：本发明提供的眼底图像增强系统的总体组成示意图；
图3：8个方向滤波器的方向示意图。
具体实施方式
为了使本发明的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本发明进行进一步详细说明。
如图1所示，本发明提供的眼底图像增强方法，包括如下步骤：
步骤A：利用眼底学习图像构建血管字典，血管字典中包括设定数量的第一子图像块。
步骤A具体包括：
步骤A1：将眼底学习图像分割成若干大小相同的第一子图像块，第一子图像块的数量 应大于设定数量。可根据眼底学习图像的人工分割结果，将眼底学习图像分割为8＊8大小的若干第一子图像块，这些第一子图像块要包含粗血管、细弱细血管、高亮部分等眼底图像特征。
步骤A2：将各第一子图像块两两进行内积。内积越小，说明两个第一子图像块的相似度越小。
步骤A3：选取内积最小的设定数量个第一子图像块构建血管字典。假设设定数量为K，选取最不相似的K个子图像块构成血管字典。
步骤B：对待增强的眼底图像进行Frangi滤波，并将Frangi滤波得到的图像划分为若干相互重叠的第二子图像块。
步骤B具体包括：
步骤B1：设待增强的眼底图像为I(x，y)，尺度为σ的二维高斯函数为G(x，y；σ)，利用二维高斯函数对待增强的眼底图像I(x，y)进行平滑处理，得到平滑图像Iσ(x，y)：
Figure PCTCN2016099026-appb-000037
其中，
Figure PCTCN2016099026-appb-000038
为卷积操作。
步骤B2：在尺度σ下，计算平滑图像Iσ(x，y)中点(x，y)处的Hessian矩阵Hσ(x，y)：
Figure PCTCN2016099026-appb-000039
步骤B3：对Hessian矩阵Hσ(x，y)做特征值分析，得到特征值λ1、λ2，|λ1|＜|λ2|。若点(x，y)属于管状结构，则|λ1|≈0，|λ2|的值会比较大，则尺度s下的血管特征为：
Figure PCTCN2016099026-appb-000040
其中，
Figure PCTCN2016099026-appb-000041
β和C是预设常数。
步骤B4：在多尺度下，去各尺度下υ0(s)的最大值作为待增强的眼底图像I(x，y)的Frangi滤波结果v：
Figure PCTCN2016099026-appb-000042
其中，smin和smax分别是最小尺度和最大尺度。
步骤B5：将Frangi滤波结果v划分为若干相互重叠的第二子图像块。
步骤C：利用方向滤波器对第二子图像块进行方向滤波，并根据方向滤波结果判断第二子图像块中包含的眼底血管是粗血管还是细弱血管。
步骤C具体包括：
步骤C1：设置方向分别为θ1＝0，
Figure PCTCN2016099026-appb-000043
Figure PCTCN2016099026-appb-000044
的8个方向滤波器(如图3所示)。
步骤C2：假设方向为θi的方向滤波器中血管区域(白色区域)为Ω1，非血管区域(黑色区域)为Ω2，计算两个区域各自的能量
Figure PCTCN2016099026-appb-000045
和
Figure PCTCN2016099026-appb-000046
Figure PCTCN2016099026-appb-000047
Figure PCTCN2016099026-appb-000048
其中v(x，y)是Frangi滤波结果v在(x，y)的值，N1是Ω1中像素个数，N2是Ω2中像素个数。
步骤C3：计算
Figure PCTCN2016099026-appb-000049
与
Figure PCTCN2016099026-appb-000050
的能量差：
Figure PCTCN2016099026-appb-000051
步骤C4：确定上述8个方向中最大的能量差：
Figure PCTCN2016099026-appb-000052
步骤C5：根据Emax判断血管类型，如果Emax≥T，则第二子图像块中包含的眼底图像为粗血管，否则为细弱血管。T为预设值。
步骤D：确定第二子图像块中的血管区域，并根据第二子图像块中包含的眼底血管的类型设置第二子图像块中的血管区域的残差权重和残差阈值。
步骤D具体包括：
步骤D1：将8个方向中最大的能量差所对应的方向滤波器(即使Emax最大的θi所对应的方向滤波器)中的血管区域作为血管实际的区域Ω1，将8个方向中最大的能量差所对应的方向滤波器(即使Emax最大的θi所对应的方向滤波器)中的非血管区域作为非血管实际的区域Ω2。
步骤D2：对于包含的眼底图像为粗血管的第二子图像块，将其血管区域Ω1的残差权重设为1，残差阈值TR＝T1；对于包含的眼底图像为细弱血管的第二子图像块，将其血管区域Ω1的残差权重设为1/υmax，残差阈值TR＝T2，其中υmax是该第二子图像块Frangi滤波结果的最大值。
步骤D还可包括：
步骤D3：设置选中的第一子图像块索引集合S为空，S＝φ，将选中的第一子图像块dr0的索引号r0加入集合S，S＝S∪r0。
步骤E：将第二子图像块与血管字典中的各第一子图像块内积，确定出其中内积最大的第一子图像块，并计算内积最大的第一子图像块对应的稀疏系数。
步骤E具体包括：
步骤E1：将第二子图像块向量化为x，血管字典中第i个第一子图像块为di。
步骤E2：将血管字典中各第一子图像块与第二子图像块x内积最大者作为选中的第一个第一子图像块dr0：
Figure PCTCN2016099026-appb-000053
其中，k为血管字典中第一子图像块的个数，r0是字典的索引号，＜x，di＞是x与di的内积运算。
步骤E3：计算第一子图像块dr0对应的稀疏系数αr0：
αr0＝＜x，dr0＞，将选中的第一子图像块dr0的索引号r0加入集合S，S＝S∪r0。
步骤F：利用内积最大的第一子图像块和第二子图像块计算残差图像，并利用所述血管区域的残差权重计算第二子图像块中血管区域的残差。
步骤F具体包括：
步骤F1：计算第二子图像块中血管区域的残差图像R：
R＝x-＜x，dr0＞dr0；
步骤F2：将残差R乘以该第二子图像块中血管区域的残差权重加权求和，作为第二子图像块中血管区域的最终残差。
步骤G：当残差的范数大于残差阈值时，将残差图像设置为第二子图像块，并跳转至步骤E，否则，跳转至步骤H。即在步骤F中的残差R的范数||R||大于残差阈值TR，则转至步骤E，否则转至步骤H。
步骤H：利用稀疏系数重构第二子图像块。重构的第二子图像块为：
Figure PCTCN2016099026-appb-000054
其中，S是多次执行步骤E确定出的多个稀疏系数的集合，dr0是每一次执行步骤E确定出的内积最大的第一子图像块，αr0是dr0对应的稀疏系数。
步骤I：利用各重构的第二子图像块重构眼底图像，从而得到增强的眼底图像。
步骤I包括：
将所有重构的第二子图像块的不相交的部分合并，得到完整的增强的眼底图像。
如图2所示，基于前述眼底图像增强方法，本发明还提供了一种眼底图像增强系统，包括：血管字典构建模块1、图像滤波及划分模块2、血管类型判断模块5、血管区域及其残差权重和残差阈值确定模块4、稀疏系数计算模块3、血管区域残差计算模块6、跳转模块7、第二子图像块重构模块8、眼底图像重构模块9。
血管字典构建模块1利用眼底学习图像构建血管字典，血管字典中包括设定数量的第一子图像块。血管字典构建模块1包括眼底学习图像划分模块、第一子图像块内积模块、血管字典构建子模块。
眼底学习图像划分模块将眼底学习图像分割成若干大小相同的第一子图像块，第一子图 像块的数量大于设定数量。
第一子图像块内积模块将各第一子图像块两两进行内积。
血管字典构建子模块选取内积最小的设定数量个第一子图像块构建血管字典。
图像滤波及划分模块2对待增强的眼底图像进行Frangi滤波，并将Frangi滤波得到的图像划分为若干相互重叠的第二子图像块。图像滤波及划分模块2包括平滑滤波模块、Hessian矩阵计算模块、特征值分析模块、Frangi滤波结果生成模块、第二子图像划分模块。
平滑滤波模块设待增强的眼底图像为I(x，y)，尺度为σ的二维高斯函数为G(x，y；σ)，利用二维高斯函数对待增强的眼底图像I(x，y)进行平滑处理，得到平滑图像Iσ(x，y)：
Figure PCTCN2016099026-appb-000055
其中，
Figure PCTCN2016099026-appb-000056
为卷积操作。
Hessian矩阵计算模块在尺度σ下，计算平滑图像Iσ(x，y)中点(x，y)处的Hessian矩阵Hσ(x，y)：
Figure PCTCN2016099026-appb-000057
特征值分析模块对Hessian矩阵Hσ(x，y)做特征值分析，得到特征值λ1、λ2，|λ1|＜|λ2|；尺度s下的血管特征为：
Figure PCTCN2016099026-appb-000058
其中，
Figure PCTCN2016099026-appb-000059
β和C是预设常数。
Frangi滤波结果生成模块取各尺度下υ0(s)的最大值作为待增强的眼底图像I(x，y)的Frangi滤波结果v：
Figure PCTCN2016099026-appb-000060
其中，smin和smax分别是最小尺度和最大尺度。
第二子图像划分模块将Frangi滤波结果v划分为若干相互重叠的第二子图像块。
血管类型判断模块5利用方向滤波器对第二子图像块进行方向滤波，并根据方向滤波结果判断该第二子图像块中包含的眼底血管是粗血管还是细弱血管。血管类型判断模块5包括方向滤波器设置模块、能量计算模块、能量差计算模块、最大能量差确定模块、血管类型判断子模块。
方向滤波器设置模块设置方向分别为θ1＝0，
Figure PCTCN2016099026-appb-000061
Figure PCTCN2016099026-appb-000062
的8个方向滤波器。
能量计算模块假设方向为θi的方向滤波器中血管区域为Ω1，非血管区域为Ω2，计算两个区域各自的能量
Figure PCTCN2016099026-appb-000063
和
Figure PCTCN2016099026-appb-000064
Figure PCTCN2016099026-appb-000065
Figure PCTCN2016099026-appb-000066
其中v(x，y)是Frangi滤波结果v在(x，y)的值，N1是Ω1中像素个数，N2是Ω2中像素个数。
能量差计算模块计算
Figure PCTCN2016099026-appb-000067
与
Figure PCTCN2016099026-appb-000068
的能量差：
Figure PCTCN2016099026-appb-000069
最大能量差确定模块确定上述8个方向中最大的能量差：
Figure PCTCN2016099026-appb-000070
血管类型判断子模块根据Emax判断血管类型，如果Emax≥T，则该第二子图像块中包含的眼底图像为粗血管，否则为细弱血管。
血管区域及其残差权重和残差阈值确定模块4确定第二子图像块中的血管区域，并根据第二子图像块中包含的眼底血管的类型设置第二子图像块中的血管区域的残差权重和残差阈值。血管区域及其残差权重和残差阈值确定模块4包括血管区域确定模块和残差权重和残差阈值确定模块。
血管区域确定模块将8个方向中最大的能量差所对应的方向滤波器中的血管区域作为血管实际的区域Ω1，将8个方向中最大的能量差所对应的方向滤波器中的血管区域作为血管实际的区域Ω2。
残差权重和残差阈值确定模块将包含的眼底图像为粗血管的第二子图像块的血管区域Ω1的残差权重设为1，残差阈值TR＝T1；将包含的眼底图像为细弱血管的第二子图像块的血管区域Ω1的残差权重设为1/υmax，残差阈值TR＝T2，其中υmax是该第二子图像块Frangi滤波结果的最大值。
稀疏系数计算模块3将第二子图像块与血管字典中的各第一子图像块内积，确定出其中内积最大的第一子图像块，并计算内积最大的第一子图像块对应的稀疏系数。稀疏系数计算模块3包括图像向量模块、第一子图像块选择模块和稀疏系数计算子模块。
图像向量模块将第二子图像块向量化为x，血管字典中第i个第一子图像块为di。
第一子图像块选择模块将血管字典中各第一子图像块与第二子图像块x内积最大者作为选中的第一个第一子图像块dr0：
Figure PCTCN2016099026-appb-000071
其中，k为血管字典中第一子图像块的个数，r0是字典的索引号，＜x，di＞是x与di的内积运算。
稀疏系数计算子模块计算第一子图像块dr0对应的稀疏系数αr0：
αr0＝＜x，dr0＞，将选中的第一子图像块dr0的索引号r0加入集合S，S＝S∪r0。
血管区域残差计算模块6利用内积最大的第一子图像块和第二子图像块计算残差图像，并利用所述血管区域的残差权重计算第二子图像块中血管区域的残差。血管区域残差计算模块6包括残差初算模块和残差加权模块。
残差初算模块计算第二子图像块中血管区域的残差图像R：
R＝x-＜x，dr0＞dr0；
残差加权模块将残差R乘以该第二子图像块中血管区域的残差权重加权求和，作为第二子图像块中血管区域的最终残差。
跳转模块7在当残差的范数大于残差阈值时，将残差图像设置为第二子图像块，并跳转至稀疏系数计算模块3，否则，跳转至第二子图像块重构模块8。
第二子图像块重构模块8利用稀疏系数重构第二子图像块。重构的第二子图像块为：
Figure PCTCN2016099026-appb-000072
其中，S是所述稀疏系数计算模块多次确定出的多个稀疏系数的集合，dr0是所述稀疏系数计算模块每一次确定出的内积最大的第一子图像块，αr0是dr0对应的稀疏系数。
眼底图像重构模块9利用各重构的第二子图像块重构眼底图像，从而得到增强的眼底图像。眼底图像重构模块9具体用于：
将所有重构的第二子图像块的不相交的部分合并，得到完整的增强的眼底图像。本系统中各模块的具体工作原理可参照前述眼底图像增强方法中的对应步骤。
以上仅为本发明的较佳实施例而已，并不用以限制本发明，凡在本发明的精神和原则之内所作的任何修改、等同替换和改进等，均应包含在本发明的保护范围之内。

Claims (18)

  1. 一种眼底图像增强方法，其特征在于，包括如下步骤：
    步骤A：利用眼底学习图像构建血管字典，所述血管字典中包括设定数量的第一子图像块；
    步骤B：对待增强的眼底图像进行Frangi滤波，并将Frangi滤波得到的图像划分为若干相互重叠的第二子图像块；
    步骤C：利用方向滤波器对所述第二子图像块进行方向滤波，并根据方向滤波结果判断所述第二子图像块中包含的眼底血管是粗血管还是细弱血管；
    步骤D：确定所述第二子图像块中的血管区域，并根据所述第二子图像块中包含的眼底血管的类型设置所述第二子图像块中的血管区域的残差权重和残差阈值；
    步骤E：将所述第二子图像块与所述血管字典中的各第一子图像块内积，确定出其中内积最大的第一子图像块，并计算所述内积最大的第一子图像块对应的稀疏系数；
    步骤F：利用所述内积最大的第一子图像块和所述第二子图像块计算残差图像，并利用所述血管区域的残差权重计算所述第二子图像块中血管区域的残差；
    步骤G：当所述残差的范数大于所述残差阈值时，将残差图像设置为第二子图像块，并跳转至步骤E，否则，跳转至步骤H；
    步骤H：利用所述稀疏系数重构所述第二子图像块；
    步骤I：利用各重构的第二子图像块重构所述眼底图像，从而得到增强的眼底图像。
  2. 如权利要求1所述的眼底图像增强方法，其特征在于，所述步骤A包括：
    步骤A1：将所述眼底学习图像分割成若干大小相同的第一子图像块；所述第一子图像块的数量大于所述设定数量；
    步骤A2：将各第一子图像块两两进行内积；
    步骤A3：选取内积最小的所述设定数量个第一子图像块构建所述血管字典。
  3. 如权利要求1所述的眼底图像增强方法，其特征在于，所述步骤B包括：
    步骤B1：设待增强的眼底图像为I(x，y)，尺度为σ的二维高斯函数为G(x，y；σ)，利用所述二维高斯函数对所述待增强的眼底图像I(x，y)进行平滑处理，得到平滑图像Iσ(x，y)：
    Figure PCTCN2016099026-appb-100001
    其中，
    Figure PCTCN2016099026-appb-100002
    Figure PCTCN2016099026-appb-100003
    为卷积操作；
    步骤B2：在尺度σ下，计算平滑图像Iσ(x，y)中点(x，y)处的Hessian矩阵Hσ(x，y)：
    Figure PCTCN2016099026-appb-100004
    步骤B3：对所述Hessian矩阵Hσ(x，y)做特征值分析，得到特征值λ1、λ2，|λ1|＜|λ2|； 尺度s下的血管特征为：
    Figure PCTCN2016099026-appb-100005
    其中，
    Figure PCTCN2016099026-appb-100006
    β和C是预设常数；
    步骤B4：在多尺度下，取各尺度下v0(s)的最大值作为所述待增强的眼底图像I(x，y)的Frangi滤波结果v：
    Figure PCTCN2016099026-appb-100007
    其中，smin和smax分别是最小尺度和最大尺度；
    步骤B5：将所述Frangi滤波结果v划分为若干相互重叠的第二子图像块。
  4. 如权利要求1所述的眼底图像增强方法，其特征在于，所述步骤C包括：
    步骤C1：设置方向分别为θ1＝0，
    Figure PCTCN2016099026-appb-100008
    Figure PCTCN2016099026-appb-100009
    的8个方向滤波器；
    步骤C2：假设方向为θi的方向滤波器中血管区域为Ω1，非血管区域为Ω2，计算两个区域各自的能量
    Figure PCTCN2016099026-appb-100010
    和
    Figure PCTCN2016099026-appb-100011
    Figure PCTCN2016099026-appb-100012
    Figure PCTCN2016099026-appb-100013
    其中v(x，y)是Frangi滤波结果v在(x，y)的值，N1是Ω1中像素个数，N2是Ω2中像素个数；
    步骤C3：计算
    Figure PCTCN2016099026-appb-100014
    与
    Figure PCTCN2016099026-appb-100015
    的能量差：
    Figure PCTCN2016099026-appb-100016
    步骤C4：确定上述8个方向中最大的能量差：
    Figure PCTCN2016099026-appb-100017
    步骤C5：根据所述Emax判断血管类型，如果Emax≥T，则所述第二子图像块中包含的眼底图像为粗血管，否则为细弱血管。
  5. 如权利要求3所述的眼底图像增强方法，其特征在于，所述步骤D包括：
    步骤D1：将所述8个方向中最大的能量差所对应的方向滤波器中的血管区域作为血管实际的区域Ω1，将所述8个方向中最大的能量差所对应的方向滤波器中的非血管区域作为非血管实际的区域Ω2；
    步骤D2：对于包含的眼底图像为粗血管的第二子图像块，将其血管区域Ω1的残差权重设为1，残差阈值TR＝T1；对于包含的眼底图像为细弱血管的第二子图像块，将其血管区域Ω1的残差权重设为1/vmax，残差阈值TR＝T2，其中vmax是该第二子图像块Frangi滤波结果的最大值。
  6. 如权利要求1所述的眼底图像增强方法，其特征在于，所述步骤E包括：
    步骤E1：将所述第二子图像块向量化为x，所述血管字典中第i个第一子图像块为di；
    步骤E2：将所述血管字典中各第一子图像块与所述第二子图像块x内积最大者作为选中的第一个第一子图像块dr0：
    Figure PCTCN2016099026-appb-100018
    其中，k为所述血管字典中第一子图像块的个数，r0是字典的索引号，<x，di>是x与di的内积运算；
    步骤E3：计算第一子图像块dr0对应的稀疏系数αr0：
    αr0＝<x，dr0>。
  7. 如权利要求1所述的眼底图像增强方法，其特征在于，所述步骤F包括：
    步骤F1：计算所述第二子图像块中血管区域的残差图像R：
    R＝x-<x，dr0>dr0；
    步骤F2：将所述残差R乘以该第二子图像块中血管区域的残差权重加权求和，作为第二子图像块中血管区域的最终残差。
  8. 如权利要求1所述的眼底图像增强方法，其特征在于，重构的所述第二子图像块为：
    Figure PCTCN2016099026-appb-100019
    其中，S是多次执行步骤E确定出的多个稀疏系数的集合，dr0是每一次执行步骤E确定出的内积最大的第一子图像块，αr0是dr0对应的稀疏系数。
  9. 如权利要求1所述的眼底图像增强方法，其特征在于，所述步骤I包括：
    将所有重构的第二子图像块的不相交的部分合并，得到完整的增强的眼底图像。
  10. 一种眼底图像增强系统，其特征在于，包括：
    血管字典构建模块，其利用眼底学习图像构建血管字典，所述血管字典中包括设定数量的第一子图像块；
    图像滤波及划分模块，其对待增强的眼底图像进行Frangi滤波，并将Frangi滤波得到的图像划分为若干相互重叠的第二子图像块；
    血管类型判断模块，其利用方向滤波器对所述第二子图像块进行方向滤波，并根据方向滤波结果判断所述第二子图像块中包含的眼底血管是粗血管还是细弱血管；
    血管区域及其残差权重和残差阈值确定模块，其确定所述第二子图像块中的血管区域，并根据所述第二子图像块中包含的眼底血管的类型设置所述第二子图像块中的血管区域的残差权重和残差阈值；
    稀疏系数计算模块，其将所述第二子图像块与所述血管字典中的各第一子图像块内积， 确定出其中内积最大的第一子图像块，并计算所述内积最大的第一子图像块对应的稀疏系数；
    血管区域残差计算模块，其利用所述内积最大的第一子图像块和所述第二子图像块计算残差图像，并利用所述血管区域的残差权重计算所述第二子图像块中血管区域的残差；
    跳转模块，其在当所述残差的范数大于所述残差阈值时，将残差图像设置为第二子图像块，并跳转至所述稀疏系数计算模块，否则，跳转至第二子图像块重构模块；
    第二子图像块重构模块，其利用所述稀疏系数重构所述第二子图像块；
    眼底图像重构模块，其利用各重构的第二子图像块重构所述眼底图像，从而得到增强的眼底图像。
  11. 如权利要求10所述的眼底图像增强系统，其特征在于，所述血管字典构建模块包括：
    眼底学习图像划分模块，其将所述眼底学习图像分割成若干大小相同的第一子图像块；所述第一子图像块的数量大于所述设定数量；
    第一子图像块内积模块，其将各第一子图像块两两进行内积；
    血管字典构建子模块，其选取内积最小的所述设定数量个第一子图像块构建所述血管字典。
  12. 如权利要求10所述的眼底图像增强系统，其特征在于，所述图像滤波及划分模块包括：
    平滑滤波模块，其设待增强的眼底图像为I(x，y)，尺度为σ的二维高斯函数为G(x，y；σ)，利用所述二维高斯函数对所述待增强的眼底图像I(x，y)进行平滑处理，得到平滑图像Iσ(x，y)：
    Figure PCTCN2016099026-appb-100020
    其中，
    Figure PCTCN2016099026-appb-100021
    Figure PCTCN2016099026-appb-100022
    为卷积操作；
    Hessian矩阵计算模块，其在尺度σ下，计算平滑图像Iσ(x，y)中点(x，y)处的Hessian矩阵Hσ(x，y)：
    Figure PCTCN2016099026-appb-100023
    特征值分析模块，其对所述Hessian矩阵Hσ(x，y)做特征值分析，得到特征值λ1、λ2，|λ1|＜|λ2|；尺度s下的血管特征为：
    Figure PCTCN2016099026-appb-100024
    其中，
    Figure PCTCN2016099026-appb-100025
    β和C是预设常数；
    Frangi滤波结果生成模块，其取各尺度下v0(s)的最大值作为所述待增强的眼底图像 I(x，y)的Frangi滤波结果v：
    Figure PCTCN2016099026-appb-100026
    其中，smin和smax分别是最小尺度和最大尺度；
    第二子图像划分模块，其将所述Frangi滤波结果v划分为若干相互重叠的第二子图像块。
  13. 如权利要求10所述的眼底图像增强系统，其特征在于，所述血管类型判断模块包括：
    方向滤波器设置模块，其设置方向分别为θ1＝0，
    Figure PCTCN2016099026-appb-100027
    Figure PCTCN2016099026-appb-100028
    的8个方向滤波器；
    能量计算模块，其假设方向为θi的方向滤波器中血管区域为Ω1，非血管区域为Ω2，计算两个区域各自的能量
    Figure PCTCN2016099026-appb-100029
    和
    Figure PCTCN2016099026-appb-100030
    Figure PCTCN2016099026-appb-100031
    Figure PCTCN2016099026-appb-100032
    其中v(x，y)是Frangi滤波结果v在(x，y)的值，N1是Ω1中像素个数，N2是Ω2中像素个数；
    能量差计算模块，其计算
    Figure PCTCN2016099026-appb-100033
    与
    Figure PCTCN2016099026-appb-100034
    的能量差：
    Figure PCTCN2016099026-appb-100035
    最大能量差确定模块，其确定上述8个方向中最大的能量差：
    Figure PCTCN2016099026-appb-100036
    血管类型判断子模块，其根据所述Emax判断血管类型，如果Emax≥T，则所述第二子图像块中包含的眼底图像为粗血管，否则为细弱血管。
  14. 如权利要求13所述的眼底图像增强系统，其特征在于，所述血管区域及其残差权重和残差阈值确定模块包括：
    血管区域确定模块，其将所述8个方向中最大的能量差所对应的方向滤波器中的血管区域作为血管实际的区域Ω1，将所述8个方向中最大的能量差所对应的方向滤波器中的血管区域作为血管实际的区域Ω2；
    残差权重和残差阈值确定模块，其将包含的眼底图像为粗血管的第二子图像块的血管区域Ω1的残差权重设为1，残差阈值TR＝T1；将包含的眼底图像为细弱血管的第二子图像块的血管区域Ω1的残差权重设为1/vmax，残差阈值TR＝T2，其中vmax是该第二子图像块Frangi滤波结果的最大值。
  15. 如权利要求10所述的眼底图像增强系统，其特征在于，所述稀疏系数计算模块包括：
    图像向量模块，其将所述第二子图像块向量化为x，所述血管字典中第i个第一子图像块为di；
    第一子图像块选择模块，其将所述血管字典中各第一子图像块与所述第二子图像块x内 积最大者作为选中的第一个第一子图像块dr0：
    Figure PCTCN2016099026-appb-100037
    其中，k为所述血管字典中第一子图像块的个数，r0是字典的索引号，<x，di>是x与di的内积运算；
    稀疏系数计算子模块，其计算第一子图像块dr0对应的稀疏系数αr0：
    αr0＝<x，dr0>。
  16. 如权利要求10所述的眼底图像增强方法，其特征在于，所述血管区域残差计算模块包括：
    残差初算模块，其计算所述第二子图像块中血管区域的残差图像R：
    R＝x-<x，dr0>dr0；
    残差加权模块，其将所述残差R乘以该第二子图像块中血管区域的残差权重加权求和，作为第二子图像块中血管区域的最终残差。
  17. 如权利要求10所述的眼底图像增强方法，其特征在于，重构的所述第二子图像块为：
    Figure PCTCN2016099026-appb-100038
    其中，S是所述稀疏系数计算模块多次确定出的多个稀疏系数的集合，dr0是所述稀疏系数计算模块每一次确定出的内积最大的第一子图像块，αr0是dr0对应的稀疏系数。
  18. 如权利要求10所述的眼底图像增强方法，其特征在于，所述眼底图像重构模块具体用于：
    将所有重构的第二子图像块的不相交的部分合并，得到完整的增强的眼底图像。
PCT/CN2016/099026 2016-09-14 2016-09-14 一种眼底图像增强方法及系统 WO2018049598A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/CN2016/099026 WO2018049598A1 (zh) 2016-09-14 2016-09-14 一种眼底图像增强方法及系统
US16/221,781 US10748268B2 (en) 2016-09-14 2018-12-17 Retinal vessel image enhancement method and system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2016/099026 WO2018049598A1 (zh) 2016-09-14 2016-09-14 一种眼底图像增强方法及系统

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US16/221,781 Continuation US10748268B2 (en) 2016-09-14 2018-12-17 Retinal vessel image enhancement method and system

Publications (1)

Publication Number Publication Date
WO2018049598A1 true WO2018049598A1 (zh) 2018-03-22

Family

ID=61619015

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2016/099026 WO2018049598A1 (zh) 2016-09-14 2016-09-14 一种眼底图像增强方法及系统

Country Status (2)

Country Link
US (1) US10748268B2 (zh)
WO (1) WO2018049598A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109685871A (zh) * 2018-12-14 2019-04-26 深圳先进技术研究院 图像重建方法、计算机设备及计算机可读存储介质
CN110111429A (zh) * 2019-03-16 2019-08-09 哈尔滨理工大学 一种检测单像素血管的方法
CN111028241A (zh) * 2019-11-13 2020-04-17 东北大学 一种多尺度血管增强的水平集分割系统与方法
CN115205835A (zh) * 2022-09-08 2022-10-18 山东特联信息科技有限公司 液化石油气标签图形识别系统

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111724335A (zh) * 2019-03-21 2020-09-29 深圳中科飞测科技有限公司 检测方法及检测系统
CN110458845A (zh) * 2019-06-25 2019-11-15 上海圭目机器人有限公司 基于图像相似度的无人机航拍图像差异分析方法
US20230025493A1 (en) * 2019-12-24 2023-01-26 Kowa Company, Ltd. Ophthalmic image processing method, ophthalmic image processing device, and ophthalmic image processing program

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20090136112A1 (en) * 2007-11-23 2009-05-28 Vincent Bismuth Method for the processing of images in interventional radioscopy
CN101980284A (zh) * 2010-10-26 2011-02-23 北京理工大学 基于两尺度稀疏表示的彩色图像降噪方法
CN104200232A (zh) * 2014-09-02 2014-12-10 西安电子科技大学 基于滑动窗融合的两次稀疏表示的图像处理方法

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP5060447B2 (ja) * 2008-10-07 2012-10-31 株式会社東芝 ノイズキャンセル処理回路および固体撮像装置
CA2938077A1 (en) * 2013-03-15 2014-09-18 Bio-Tree Systems, Inc. Methods and system for linking geometry obtained from images
US10402965B1 (en) * 2015-11-12 2019-09-03 Carl Zeiss Meditec, Inc. Systems and methods for reducing artifacts in OCT angiography images

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20090136112A1 (en) * 2007-11-23 2009-05-28 Vincent Bismuth Method for the processing of images in interventional radioscopy
CN101980284A (zh) * 2010-10-26 2011-02-23 北京理工大学 基于两尺度稀疏表示的彩色图像降噪方法
CN104200232A (zh) * 2014-09-02 2014-12-10 西安电子科技大学 基于滑动窗融合的两次稀疏表示的图像处理方法

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109685871A (zh) * 2018-12-14 2019-04-26 深圳先进技术研究院 图像重建方法、计算机设备及计算机可读存储介质
WO2020118894A1 (zh) * 2018-12-14 2020-06-18 深圳先进技术研究院 图像重建方法、计算机设备及计算机可读存储介质
CN109685871B (zh) * 2018-12-14 2020-07-24 深圳先进技术研究院 图像重建方法、计算机设备及计算机可读存储介质
CN110111429A (zh) * 2019-03-16 2019-08-09 哈尔滨理工大学 一种检测单像素血管的方法
CN110111429B (zh) * 2019-03-16 2022-11-18 哈尔滨理工大学 一种检测单像素血管的方法
CN111028241A (zh) * 2019-11-13 2020-04-17 东北大学 一种多尺度血管增强的水平集分割系统与方法
CN111028241B (zh) * 2019-11-13 2023-10-10 东北大学 一种多尺度血管增强的水平集分割系统与方法
CN115205835A (zh) * 2022-09-08 2022-10-18 山东特联信息科技有限公司 液化石油气标签图形识别系统

Also Published As

Publication number Publication date
US10748268B2 (en) 2020-08-18
US20190122350A1 (en) 2019-04-25

Similar Documents

Publication Publication Date Title
WO2018049598A1 (zh) 一种眼底图像增强方法及系统
CN110706246B (zh) 一种血管图像分割方法、装置、电子设备和存储介质
CN109461495A (zh) 一种医学图像的识别方法、模型训练的方法及服务器
Chen et al. 3D intracranial artery segmentation using a convolutional autoencoder
Wang et al. Frnet: an end-to-end feature refinement neural network for medical image segmentation
Guo et al. Retinal vessel segmentation combined with generative adversarial networks and dense U-Net
CN115205300B (zh) 基于空洞卷积和语义融合的眼底血管图像分割方法与系统
CN113205524B (zh) 基于U-Net的血管图像分割方法、装置和设备
Manakov et al. Noise as domain shift: Denoising medical images by unpaired image translation
CN106504199B (zh) 一种眼底图像增强方法及系统
CN111583262A (zh) 一种血管分割方法与系统
CN113298830B (zh) 一种基于自监督的急性颅内ich区域图像分割方法
CN109003280B (zh) 一种双通道血管内超声影像的血管中内膜分割方法
CN113920078A (zh) 一种针对眼底图像的血管分割方法和电子设备
Moghari et al. Efficient radiation dose reduction in whole-brain CT perfusion imaging using a 3D GAN: performance and clinical feasibility
EP4315245A2 (en) Medical image analysis using neural networks
CN115830163A (zh) 基于深度学习的确定性引导的渐进式医学图像跨模态生成方法和装置
CN113781403A (zh) 一种胸部ct影像处理方法与装置
Bhuvaneswari et al. Contrast enhancement of retinal images using green plan masking and whale optimization algorithm
CN116645283A (zh) 基于自监督感知损失多尺度卷积神经网络的低剂量ct图像去噪方法
Beevi et al. Denoising transthoracic echocardiographic images in regional wall motion abnormality using deep learning techniques
Khan et al. A Computer-Aided Diagnostic System to Identify Diabetic Retinopathy, Utilizing a Modified Compact Convolutional Transformer and Low-Resolution Images to Reduce Computation Time. Biomedicines. 2023. No. 11. Art. 1566
Akkasaligar et al. Analysis of polycystic kidney disease in medical ultrasound images
Zhang et al. Blood vessel segmentation based on digital subtraction angiography sequence
Arpacı et al. LUPU-Net: a new improvement proposal for encoder-decoder architecture

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 16915983

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 10/07/2019)

122 Ep: pct application non-entry in european phase

Ref document number: 16915983

Country of ref document: EP

Kind code of ref document: A1